Pith. sign in

REVIEW 2 major objections 7 minor 20 references

Pitwall: Faithful Natural-Language Race-Strategy Briefings from a Calibrated Real-Time Monte Carlo Engine

T0 review · 2 major / 7 minor · reviewed 2026-07-08 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read F1 strategy engine generates verified trilingual race briefings

desk verdict Private letter on arXiv:2607.06495 read the letter →

arxiv 2607.06495 v1 pith:INZVH4I5 submitted 2026-07-07 cs.CL cs.AIcs.LGstat.AP

Juan S. Santillana (Independent Researcher) This is my paper
classification cs.CLcs.AIcs.LGstat.AP
keywords Formula1MonteCarlosimulationcalibrationgroundedgenerationfaithfulnessracestrategynaturallanguageBrierscore
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper presents Pitwall, a production system that produces natural-language Formula 1 race-strategy briefings in English, Spanish, and Portuguese, where every published sentence is decomposed into typed factual claims and verified against a probabilistic race state before release. The grounding substrate is a vectorized Monte Carlo engine that simulates 2,000 race continuations per lap, calibrated on 126 races (2018-2024) and validated on fully held-out 2025-2026 seasons with a Brier score of 0.0745. The central methodological argument is that desirable system properties—realism, calibration, decision quality, and generation richness—are distinct objectives that repeatedly trade off against each other and must therefore be gated separately. In simulation, a learned overtake kernel that improves pass prediction does not improve field-level calibration; a compound-pace normalization that fixes pathological pit recommendations worsens outcome prediction. In generation, fine-tuning on richer verified targets produces more vivid text but causes hallucination when the grounding state is sparse—a failure traced across four base models to instruction adherence rather than scale. The system routes each component to the path it improves: a Brier-optimal configuration for calibrated probabilities and a feasibility-aware configuration for recommendations. End-to-end live operation was confirmed at two consecutive 2026 Grands Prix, including a pre-committed probability trace that locked onto the eventual winner ten laps before the finish.

What carries the argument

The central mechanism is verifier-gated generation: every published sentence is decomposed into typed factual claims (positions, gaps, tyres, pace, overtakes, race control) and each claim is checked against the live probabilistic race state. The same verifier gates the fine-tuning corpus—of 3,045 model-written targets, only the 81.9% with zero contradicted claims are retained, the rest fall back to a provably faithful template, so the generator never trains on an ungrounded example. The grounding substrate is a vectorized Monte Carlo engine using common random numbers for variance-controlled undercut counterfactuals, with a dual-path architecture routing Brier-optimal components to the path.

What would settle it

If the system's undercut window probabilities were systematically miscalibrated on races with high undercut activity (street circuits, close battles), this would indicate the open-loop rival policy is insufficient and the covering game matters.

Watch

Extended reading notes

Core claim

The paper's core finding is that in a live decision-support system combining simulation and natural-language generation, the components that improve task-level accuracy or realism are not the ones that improve calibrated probability estimates or faithful text generation, and that these objectives actively conflict. A learned overtake model that beats a baseline on pass prediction yields no field-calibration gain after recalibration. A compound-pace projection that eliminates dangerous pit-call recommendations degrades Brier score because the confounded raw fit encodes genuine race-specific signal. A language model fine-tuned on faithfulness-gated targets hallucinates precisely when the input

Load-bearing premise

The pace-optimal open-loop rival policy assumes all 19 rival cars execute their deterministic pace-optimal remaining strategy and do not react to the focal car's pit decisions. If this is substantially wrong—real teams do react to undercut attempts—then the undercut window probabilities and strategy recommendations are systematically biased.

Editorial extensions

If this is right

  • The dual-path principle—routing calibration-optimal and decision-optimal components separately rather than shipping one model everywhere—generalizes to any decision-support system where a single model serves both probability estimation and recommendation generation.
  • The finding that faithfulness-gated fine-tuning is necessary but not sufficient, and that sparse-context hallucination tracks base-model instruction adherence rather than scale, suggests that base-model selection under sparse-context audit should be a standard gate in any grounded generation pipeline.
  • The parity-gating discipline—where every new component ships behind a flag whose default reproduces the incumbent engine bit-for-bit—provides a template for incrementally adding complexity to a calibrated simulator without silent behavioral drift.
  • The negative result that a learned overtake kernel improves task-level prediction but not field-level calibration illustrates how calibration optimizers can absorb the effect of a better sub-model by adjusting other parameters, making the binding constraint on outcome fidelity something other than the sub-model's accuracy.
  • Live deployment surfaced two distinct failure families—state estimation (out-lap contamination) and prior selection (wet-regime inheritance)—that no backtest exercises, suggesting that live operation is an irreplaceable test for real-time systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The calibration/decision split may be a specific instance of a more general phenomenon: any time a confounded model is re-calibrated, the confound may carry genuine signal for the confounded task while producing pathological outputs for a related task, and the two uses cannot be served by a single de-confounded model.
  • The finding that sparse-context hallucination is base-model-dependent and not scale-dependent implies that the industry trend toward larger models may not resolve faithfulness issues in grounded generation without explicit instruction-adherence auditing.
  • The common-random-numbers undercut counterfactual could be extended to other real-time strategic decisions in sports (e.g., power-play timing in hockey, substitution timing in basketball) where the decision is 'act now vs. wait' and the variance of the contrast matters more than the variance of either option alone.
  • The analog-prior spread guardrail (rejecting a prior when fitted compound spreads exceed physical plausibility) suggests a general principle: domain-specific physical constraints should audit statistical priors before they enter a live system, not just after they produce bad outputs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. Pitwall is a production system for generating faithful, trilingual (English/Spanish/Portuguese) Formula 1 race-strategy briefings from a calibrated, real-time Monte Carlo engine. The system ingests the official F1 live-timing stream, maintains a canonical probabilistic race state, evaluates strategic counterfactuals (e.g., undercut windows) via vectorized Monte Carlo simulation, and verbalizes the results using a fine-tuned language model. The paper makes two primary contributions: (1) a calibrated simulation engine with a rigorous gating protocol that separates an 'oracle path' (for calibrated outcome probabilities) from a 'decision path' (for strategy recommendations), and (2) a verifier-gated language layer where every published sentence is decomposed into typed factual claims and verified against the race state. The engine is calibrated on 126 races (2018-2024) and validated on fully held-out 2025-2026 seasons (Brier 0.0745). End-to-end operation was confirmed at two live Grands Prix (Austria and Britain, 2026).

Significance. The paper makes significant contributions to both sports analytics and grounded natural language generation. The dual-path principle (Section 5.3)—routing components to either an oracle path or a decision path based on held-out evidence—is a clean, generalizable methodological insight, supported by honest negative results (e.g., the H7 overtake kernel improving task-level prediction but not field-level calibration, Table 4). The verifier-gated fine-tuning approach (Section 6.2), where the faithfulness metric gates the training data itself rather than just evaluating post-hoc, is a novel and practical contribution to reference-free grounded generation. The system ships with reproducible code, calibrated parameters with physically interpretable bounds (Table 2), and falsifiable predictions validated on fully held-out seasons. The four-base replication isolating instruction adherence as the driver of sparse-context hallucination (Section 7.7) is a valuable empirical finding for the NLP community.

major comments (2)
  1. Section 7.4: The H14 undercut window is described as 'the core live decision' (Section 4.7), producing probabilities P̂_h(ahead) and deltas Δ̂_h that form the direct basis for strategy recommendations. However, these MC engine undercut probabilities are never directly calibrated against realized undercut outcomes. Section 7.4 extracts 2,134 undercut events and fits a *separate* logistic regression model (Brier 0.168 on 828 held-out events), which the paper explicitly states 'validates feature importance'—not the MC engine's own undercut probabilities. This is a load-bearing gap: the central claim is that Pitwall produces faithful briefings from a *calibrated* MC engine, but the engine's actionable decision output (H14) lacks direct held-out calibration against ground truth. The overall finish-position Brier (0.0745) could remain good while the undercut window probabilities are miscaliber
  2. Section 4.6, H1: The pace-optimal open-loop rival policy assumes all 19 rival cars execute their deterministic pace-optimal remaining strategy and do not react to the focal car's pit decisions. The paper acknowledges this as a limitation (Section 8(v)), identifying the covering game as the 'main residual error source in undercut-heavy moments.' This is not merely a known limitation—it directly interacts with the H14 calibration gap above. If real rivals cover the undercut, the engine's P̂_h(ahead) for box-now would be systematically inflated relative to realized swap rates. The paper should either (a) provide a sensitivity analysis showing how large the rival-reactivity bias would need to be to flip H14 recommendations on marginal calls, or (b) explicitly scope the 'calibrated' claim to finish-position probabilities (win/podium/points) and acknowledge that the undercut window is variance
minor comments (7)
  1. Table 1: The 'parity-gated' footnote (asterisk) for H15, H17, H18 states these are 'bit-identical to the base engine at default settings,' but the table also lists H17 as serving the 'decision' path. Clarify whether H17 is active on the decision path or still pending its gating experiment.
  2. Section 7.4: The logistic regression's tyre-age delta coefficient (ŵ=+0.01) is described as 'positive but weak.' Given the base rate of 25% and N=828, a coefficient of 0.01 on a presumably standardized input seems negligibly small—consider reporting whether this is statistically significant or effectively zero.
  3. Figure 2: The reliability diagrams are informative but the bin counts (bar underlay) are difficult to read in the 0.8-0.9 win bin where 'mild over-confidence' is noted. Consider adding bin counts as text annotations or a supplementary table.
  4. Section 7.6, Table 6: The live probability trace covers only 14 polls from one race. The paper correctly notes this is 'an anecdote, not calibration evidence,' but the abstract's claim that the system 'locked onto the eventual winner ten laps before the flag' could be read as stronger evidence than it is. Consider softening the abstract language.
  5. Section 6.1: The compound vocabulary initially omitted hypersoft, ultrasoft, and supersoft compounds (~150 occurrences). This is a useful finding but the paper does not state whether these compounds are now fully supported in the production verifier. Confirm.
  6. Equation (3): The notation for the projection uses both 'c' (compound vector) and 'c_t' (compound base pace). Using 'b' for the fitted bases and 'c' for the canonical centroid is clear, but the subscript 't' in 'c_t' is ambiguous—clarify whether this indexes compound type or time.
  7. Section 7.7: The four-base replication is described as a 'qualitative audit' rather than a 'large-scale benchmark.' The sample size (number of probe contexts) is not stated. Report the number of probes per condition to allow readers to assess the reliability of the cross-model comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

The referee raises two interconnected and correct points about the calibration status of the H14 undercut window. We agree with both and will revise accordingly.

read point-by-point responses

  1. Referee: Section 7.4: H14 undercut probabilities are never directly calibrated against realized undercut outcomes. The logistic regression in Section 7.4 validates feature importance, not the MC engine's own P̂_h(ahead). This is a load-bearing gap because the central claim is faithful briefings from a calibrated MC engine, but the engine's actionable decision output lacks direct held-out calibration.

    Authors: The referee is correct. The paper as written conflates two distinct calibration claims: (1) the oracle-path finish-position probabilities (win/podium/points, Brier 0.0745) are directly calibrated against realized outcomes on held-out seasons, and (2) the decision-path H14 undercut probabilities P̂_h(ahead) are variance-controlled via common random numbers but are not directly validated against realized undercut swap rates. The Section 7.4 logistic regression (Brier 0.168 on 828 held-out events) demonstrates that undercut outcomes are statistically predictable from pre-stop features and establishes gap_before as the dominant input—it does not calibrate the MC engine's own P̂_h(ahead). We will revise the manuscript to make this distinction explicit. Specifically: (a) we will scope the term 'calibrated' throughout to refer to oracle-path finish-position probabilities and add a clear statement that H14 undercut probabilities are CRN-stabilized contrast estimates, not directly calibrated against ground truth; (b) we will add a paragraph in Section 7.4 acknowledging this as a named limitation alongside the existing limitations in Section 8; and (c) we will reframe the Section 7.4 logistic model honestly as what it is—an independent feature-importance study that motivates the H9 correction, not a validation of the engine's undercut probabilities. We agree this is load-bearing for the paper's framing and the revision is necessary. revision: yes

  2. Referee: Section 4.6, H1: The open-loop rival policy assumes rivals don't react to the focal car's pit decisions. This directly interacts with the H14 calibration gap—if real rivals cover the undercut, P̂_h(ahead) for box-now would be systematically inflated. The paper should either (a) provide a sensitivity analysis showing how large the rival-reactivity bias would need to be to flip H14 recommendations on marginal calls, or (b) explicitly scope the 'calibrated' claim to finish-position probabilities and acknowledge the undercut window is variance-sensitive.

    Authors: The referee correctly identifies the interaction between the H1 open-loop assumption and the H14 calibration gap. If rivals cover the undercut in reality, the engine's P̂_h(ahead) for box-now is systematically optimistic because the simulated rival does not defend. We acknowledge this direction of bias is real and is already flagged in Section 8(v) as 'the main residual error source in undercut-heavy moments,' but the paper does not quantify its impact on marginal recommendations. We will adopt option (b) in full—scoping the calibration claim to finish-position probabilities—and will additionally provide a partial version of option (a). Specifically: we will add a sensitivity argument showing that on the marginal calls where the recommendation could flip (|Δ̂_h| < 2pp, which in our historical replays accounts for roughly 15-20% of undercut-window queries), a rival covering within one lap would need to reduce the realized swap rate by approximately 8-12 percentage points relative to the open-loop estimate to flip the recommendation. This is plausible in covering-game scenarios and confirms that marginal H14 calls should carry an explicit confidence gate noting the open-loop assumption. We will add this gate to the H14 reporting protocol description in Section 4.7 and note it as a limitation. We cannot at this time provide a full covering-game simulation (that is the sequential-game extension of Section 8(v) and H17 is only the SC-reactive first step), so the sensitivity analysis will be approximate rather than a closed-form bound. revision: partial

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper's central claims are grounded against external benchmarks (held-out Brier scores, realized race outcomes) and the one self-citation is explicitly scoped to the language-layer evaluation methodology, not the engine calibration or undercut window.

full rationale

The paper's core derivation chain is self-contained against external benchmarks. The MC engine is calibrated on 126 training races (2018-2024) and validated on fully held-out 2025-2026 seasons (Brier 0.0745, Table 4). The calibration objective (Eq. 7) uses realized outcomes y_i, not fitted values. The H14 undercut window (Eq. 4) produces probabilities from Monte Carlo simulation under common random numbers — these are not fitted to undercut outcomes and then presented as predictions of those same outcomes. Section 7.4 fits a *separate* logistic regression on undercut events to validate feature importance (gap_before as dominant predictor), which the paper explicitly distinguishes from the MC engine's own undercut probabilities: 'this logistic evaluation validates feature importance; the inter-option deltas from the MC engine are separately variance-controlled by the shared CRN seed.' The skeptic's concern that H14 probabilities are never directly calibrated against realized undercut outcomes is a correctness/coverage gap, not a circularity — the paper does not claim H14 is calibrated against undercut outcomes and then use that calibration as evidence of the same. The one self-citation (Santillana 2026) supports the claim-extraction-and-verification methodology for the language layer; it is not load-bearing for the engine calibration, the undercut window, or the faithfulness gating of fine-tuning data (which is independently verified against the live state). The faithfulness-gated fine-tuning (Section 6.2) admits targets based on verification against the race state, not against the model's own outputs — the gate is external to the generator. No step in the derivation chain reduces to its inputs by construction.

Assumptions & free parameters 20 free parameters · 6 assumptions · 0 invented entities

The system introduces no new physical entities, particles, forces, or dimensions. All modeled quantities (Safety Car rates, tyre degradation, pace offsets) correspond to observable F1 phenomena. The free parameters are all calibrated against historical race data with physically interpretable bounds. The calibration objective weights (w_B=2, w_E=1) are the most ad-hoc choices, as no sensitivity analysis is provided. The H1 open-loop rival policy is the most consequential domain assumption.

free parameters (20)
  • lap_noise = 0.18 (default)
    Per-lap pace noise in seconds; calibrated parameter in the 12-dimensional core vector
  • dnf_prob = 0.07 (default)
    Per-car retirement probability; calibrated parameter
  • sc_pace_mult = 1.40 (default)
    Lap time multiplier under Safety Car; calibrated parameter
  • sc_pit_factor = 0.45 (default)
    Pit loss multiplier under SC; calibrated parameter
  • dirty_air_window = 1.2 (default)
    Interaction gap window in seconds; calibrated parameter
  • dirty_air_loss = 0.35 (default)
    Held-up loss in s/lap; calibrated parameter
  • pace_clip = 0.6 (default)
    Cap on driver pace offset; calibrated parameter, noted as persistently pinned at upper bound
  • pace_scale = 0.55 (default)
    Compression of raw pace deltas; calibrated parameter, noted as persistently pinned at upper bound
  • grid_spread = 0.28 (default)
    Start spread per grid slot in seconds; calibrated parameter, noted as persistently pinned at upper bound
  • fuel_effect = 0.05 (default)
    Fuel burn-off folded out of degradation in s/lap; calibrated parameter
  • sc_rate_mult = 1.0 (default)
    Multiplier on per-circuit SC prior; calibrated parameter
  • ot_rate_mult = 1.0 (default)
    Multiplier on overtake difficulty; calibrated parameter
  • vsc_pace_mult = 1.22 (default)
    Lap time multiplier under VSC; calibrated parameter
  • vsc_pit_factor = 0.65 (default)
    Pit loss multiplier under VSC; calibrated parameter
  • vsc_dur_max = 3.0 (default)
    Max VSC duration in laps; calibrated parameter
  • w_B = 2
    Weight on Brier score in calibration objective J(theta); chosen by the author
  • w_E = 1
    Weight on ECE in calibration objective J(theta); chosen by the author
  • N_sims (calibration) = 400
    Number of MC replications per race during calibration; chosen for compute budget
  • N_sims (production) = 2000
    Number of MC replications per race in production; chosen for precision
  • B (bootstrap) = 500
    Number of bootstrap resamples for confidence intervals; chosen for compute budget
assumptions (6)
  • domain assumption Rivals execute pace-optimal remaining strategy (H1) and do not react to the focal car's pit decisions
    Section 4.6: the open-loop rival policy is a game-theoretically motivated assumption; the paper acknowledges the covering game extension as future work
  • domain assumption Tyre degradation is adequately modeled as linear wear with an optional cliff knot, plus a fuel-burn correction
    Section 4.2, Eq. 1: the lap-time model assumes a specific functional form for degradation; thermal effects are partially absorbed into noise
  • ad hoc to paper The calibration objective J(theta) = Spearman - w_B*Brier - w_E*ECE with w_B=2, w_E=1 is the right objective for a strategy engine
    Section 5.2, Eq. 7: the blending weights are chosen by the author; no sensitivity analysis on the weights is provided
  • domain assumption The public timing stream is a sufficient information source for real-time state reconstruction
    Section 8 limitation (i): no car telemetry; the system conditions on what the public feed carries
  • domain assumption Per-race compound base pace fitting on green laps yields meaningful pace estimates
    Section 4.2: green-lap fitting excludes SC/VSC, in-laps, out-laps; the fitted bases are the substrate for all downstream simulation
  • standard math The claim-extraction-and-verification methodology from Santillana [2026] is valid for live state
    Section 6.1: the verifier is extended from post-race ground truth to live state; the companion paper establishes the methodology

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pitwall: Faithful Natural-Language Race-Strategy Briefings from a Calibrated Real-Time Monte Carlo Engine." pith.science (2026). https://pith.science/paper/INZVH4I5

@misc{pith2026260706495,
  author       = {Pith},
  title        = {Pith review of: Pitwall: Faithful Natural-Language Race-Strategy Briefings from a Calibrated Real-Time Monte Carlo Engine},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INZVH4I5}},
  note         = {Machine review of arXiv:2607.06495}
}
read the original abstract

Live sports commentary is grounded generation under a deadline: statements concern real, named athletes, the grounding state changes every few seconds, and no reference text exists at generation time. We present Pitwall, a production system that generates natural-language Formula 1 strategy briefings in English, Spanish, and Portuguese, treating faithfulness as an architectural property rather than an aspiration: every published sentence is decomposed into typed factual claims (positions, gaps, tyres, pace, overtakes, race control) and each claim is verified against the probabilistic race state that prompted it. The same verifier gates the fine-tuning data: of 3,045 model-written targets, only the 81.9% whose every claim is state-supported are retained, the rest falling back to a provably faithful template, so the generator never sees an ungrounded target. Verification is meaningful because of the grounding substrate: a vectorized Monte Carlo engine (N=2,000 per-lap race continuations) calibrated on 126 races (2018-2024) and validated on fully held-out 2025-2026 seasons (winner-in-top-3 90.3% over 155 backtests; held-out Brier 0.0745). A recurring finding spans both halves of the system: virtues trade off and must be gated separately. In simulation, calibration-optimal is not decision-optimal; in generation, fine-tuning on richer targets buys vividness that collapses into hallucination when the grounding state is sparse -- a failure a four-base replication traces to base-model instruction adherence, not scale, and that sparse-context auditing removes from the production model. End-to-end operation -- live timing to verified trilingual briefings -- was confirmed at two consecutive live Grands Prix (Austria and Britain, 2026); at Silverstone a timestamped probability trace, committed to disk before the outcome was known, locked onto the eventual winner ten laps before the flag.

Figures

Figures reproduced from arXiv: 2607.06495 by the authors.

Figure 1
Figure 1. Pitwall architecture. The same structured race store (157 races, 2018–2026, excl. 2022) serves as calibration corpus, held-out backtest oracle, and grounding source for the language layer. The MC Engine runs two parallel paths: an oracle path (Brier-optimal configuration) for calibrated probabilities and a decision path (H5+H9 enabled) for recommendations. low-latency proxied edge agent and decoded into the identica… view at source ↗
Figure 2
Figure 2. Reliability diagrams on the held-out 2025–2026 test races (31 races, [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Bekker and W

    J. Bekker and W. Lotz. Planning Formula One race strategies using discrete-event simulation. Journal of the Operational Research Society, 60(7):952--961, 2009

    work page 2009

  2. [2]

    G. W. Brier. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1--3, 1950

    work page 1950

  3. [3]

    Guerrero and R

    A. Guerrero and R. Castilla. Aerodynamic study of the wake effects on a F ormula 1 car. Energies, 13(19):5183, 2020. https://doi.org/10.3390/en13195183

    work page doi:10.3390/en13195183 2020

  4. [4]

    Efron and R

    B. Efron and R. J. Tibshirani. An Introduction to the Bootstrap. Chapman & Hall/CRC, 1994

    work page 1994

  5. [5]

    Glasserman and D

    P. Glasserman and D. D. Yao. Some guidelines and guarantees for common random numbers. Management Science, 38(6):884--908, 1992

    work page 1992

  6. [6]

    Gneiting, F

    T. Gneiting, F. Balabdaoui, and A. E. Raftery. Probabilistic forecasts, calibration and sharpness. Journal of the Royal Statistical Society: Series B, 69(2):243--268, 2007

    work page 2007

  7. [7]

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. On calibration of modern neural networks. In Proceedings of ICML, 2017

    work page 2017

  8. [8]

    Hansen and A

    N. Hansen and A. Ostermeier. Completely derandomized self-adaptation in evolution strategies. Evolutionary Computation, 9(2):159--195, 2001

    work page 2001

Show all 20 references
  1. [9]

    Heilmeier, M

    A. Heilmeier, M. Graf, J. Betz, and M. Lienkamp. Application of Monte Carlo methods to consider probabilistic effects in a race simulation for circuit motorsport. Applied Sciences, 10(12):4229, 2020

    2020

  2. [10]

    Heilmeier, A

    A. Heilmeier, A. Thomaser, M. Graf, and J. Betz. Virtual strategy engineer: Using artificial neural networks for making race strategy decisions in circuit motorsport. Applied Sciences, 10(21):7805, 2020

    2020

  3. [11]

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. J. Bang, A. Madotto, and P. Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1--38, 2023

    2023

  4. [12]

    D. P. Kelly and R. S. Sharp. Time-optimal control of the race car: influence of a thermodynamic tyre model. Vehicle System Dynamics, 50(4):641--662, 2012

    2012

  5. [13]

    A. M. Law. Simulation Modeling and Analysis. McGraw-Hill, 5th edition, 2015

    2015

  6. [14]

    Maynez, S

    J. Maynez, S. Narayan, B. Bohnet, and R. McDonald. On faithfulness and factuality in abstractive summarization. In Proceedings of ACL, pages 1906--1919, 2020

    1906

  7. [15]

    A. H. Murphy. A new vector partition of the probability score. Journal of Applied Meteorology, 12(4):595--600, 1973

    1973

  8. [16]

    D. J. N. Limebeer, G. Perantoni, and A. V. Rao. Optimal control of F ormula O ne car energy recovery systems. International Journal of Control, 87(10):2065--2080, 2014. https://doi.org/10.1080/00207179.2014.900705

    doi

  9. [17]

    de Groote

    J. de Groote. Overtaking in F ormula 1 during the P irelli era: A driver-level analysis. Journal of Sports Analytics, 7(2):119--137, 2021. https://doi.org/10.3233/JSA-200466

    2021 doi

  10. [18]

    J. S. Santillana. Precision is not faithfulness: Coverage-aware evaluation of grounded generation with a complete oracle. arXiv preprint arXiv:2606.09376, 2026

    2026 arXiv

  11. [19]

    Thomson and E

    C. Thomson and E. Reiter. A gold standard methodology for evaluating accuracy in data-to-text systems. In Proceedings of INLG, pages 158--168, 2020

    2020

  12. [20]

    Wiseman, S

    S. Wiseman, S. M. Shieber, and A. M. Rush. Challenges in data-to-document generation. In Proceedings of EMNLP, pages 2253--2263, 2017

    2017

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.